Spaces:
hlby
/
Runtime error

gpt / gradio /queueing.py
hlby's picture
Duplicate from jumei/gptsite
947e9b9
from __future__ import annotations
import asyncio
import copy
import sys
import time
from asyncio import TimeoutError as AsyncTimeOutError
from collections import deque
from typing import Any, Deque, Dict, List, Tuple
import fastapi
import httpx
from gradio.data_classes import Estimation, PredictBody, Progress, ProgressUnit
from gradio.helpers import TrackedIterable
from gradio.utils import AsyncRequest, run_coro_in_background, set_task_name
class Event:
def __init__(
self,
websocket: fastapi.WebSocket,
session_hash: str,
fn_index: int,
):
self.websocket = websocket
self.session_hash: str = session_hash
self.fn_index: int = fn_index
self._id = f"{self.session_hash}_{self.fn_index}"
self.data: PredictBody | None = None
self.lost_connection_time: float | None = None
self.token: str | None = None
self.progress: Progress | None = None
self.progress_pending: bool = False
async def disconnect(self, code: int = 1000):
await self.websocket.close(code=code)
class Queue:
def __init__(
self,
live_updates: bool,
concurrency_count: int,
update_intervals: float,
max_size: int | None,
blocks_dependencies: List,
):
self.event_queue: Deque[Event] = deque()
self.events_pending_reconnection = []
self.stopped = False
self.max_thread_count = concurrency_count
self.update_intervals = update_intervals
self.active_jobs: List[None | List[Event]] = [None] * concurrency_count
self.delete_lock = asyncio.Lock()
self.server_path = None
self.duration_history_total = 0
self.duration_history_count = 0
self.avg_process_time = 0
self.avg_concurrent_process_time = None
self.queue_duration = 1
self.live_updates = live_updates
self.sleep_when_free = 0.05
self.progress_update_sleep_when_free = 0.1
self.max_size = max_size
self.blocks_dependencies = blocks_dependencies
self.access_token = ""
self.queue_client = None
async def start(self, progress_tracking=False):
# So that the client is attached to the running event loop
self.queue_client = httpx.AsyncClient()
run_coro_in_background(self.start_processing)
if progress_tracking:
run_coro_in_background(self.start_progress_tracking)
if not self.live_updates:
run_coro_in_background(self.notify_clients)
def close(self):
self.stopped = True
def resume(self):
self.stopped = False
def set_url(self, url: str):
self.server_path = url
def set_access_token(self, token: str):
self.access_token = token
def get_active_worker_count(self) -> int:
count = 0
for worker in self.active_jobs:
if worker is not None:
count += 1
return count
def get_events_in_batch(self) -> Tuple[List[Event] | None, bool]:
if not (self.event_queue):
return None, False
first_event = self.event_queue.popleft()
events = [first_event]
event_fn_index = first_event.fn_index
batch = self.blocks_dependencies[event_fn_index]["batch"]
if batch:
batch_size = self.blocks_dependencies[event_fn_index]["max_batch_size"]
rest_of_batch = [
event for event in self.event_queue if event.fn_index == event_fn_index
][: batch_size - 1]
events.extend(rest_of_batch)
[self.event_queue.remove(event) for event in rest_of_batch]
return events, batch
async def start_processing(self) -> None:
while not self.stopped:
if not self.event_queue:
await asyncio.sleep(self.sleep_when_free)
continue
if not (None in self.active_jobs):
await asyncio.sleep(self.sleep_when_free)
continue
# Using mutex to avoid editing a list in use
async with self.delete_lock:
events, batch = self.get_events_in_batch()
if events:
self.active_jobs[self.active_jobs.index(None)] = events
task = run_coro_in_background(self.process_events, events, batch)
run_coro_in_background(self.broadcast_live_estimations)
set_task_name(task, events[0].session_hash, events[0].fn_index, batch)
async def start_progress_tracking(self) -> None:
while not self.stopped:
if not any(self.active_jobs):
await asyncio.sleep(self.progress_update_sleep_when_free)
continue
for job in self.active_jobs:
if job is None:
continue
for event in job:
if event.progress_pending and event.progress:
event.progress_pending = False
client_awake = await self.send_message(
event, event.progress.dict()
)
if not client_awake:
await self.clean_event(event)
await asyncio.sleep(self.progress_update_sleep_when_free)
def set_progress(
self,
event_id: str,
iterables: List[TrackedIterable] | None,
):
if iterables is None:
return
for job in self.active_jobs:
if job is None:
continue
for evt in job:
if evt._id == event_id:
progress_data: List[ProgressUnit] = []
for iterable in iterables:
progress_unit = ProgressUnit(
index=iterable.index,
length=iterable.length,
unit=iterable.unit,
progress=iterable.progress,
desc=iterable.desc,
)
progress_data.append(progress_unit)
evt.progress = Progress(progress_data=progress_data)
evt.progress_pending = True
def push(self, event: Event) -> int | None:
"""
Add event to queue, or return None if Queue is full
Parameters:
event: Event to add to Queue
Returns:
rank of submitted Event
"""
queue_len = len(self.event_queue)
if self.max_size is not None and queue_len >= self.max_size:
return None
self.event_queue.append(event)
return queue_len
async def clean_event(self, event: Event) -> None:
if event in self.event_queue:
async with self.delete_lock:
self.event_queue.remove(event)
async def broadcast_live_estimations(self) -> None:
"""
Runs 2 functions sequentially instead of concurrently. Otherwise dced clients are tried to get deleted twice.
"""
if self.live_updates:
await self.broadcast_estimations()
async def gather_event_data(self, event: Event, receive_timeout=60) -> bool:
"""
Gather data for the event
Parameters:
event:
"""
if not event.data:
client_awake = await self.send_message(event, {"msg": "send_data"})
if not client_awake:
return False
data, client_awake = await self.get_message(event, timeout=receive_timeout)
if not client_awake:
# In the event, we timeout due to large data size
# Let the client know, otherwise will hang
await self.send_message(
event,
{
"msg": "process_completed",
"output": {"error": "Time out uploading data to server"},
"success": False,
},
)
return False
event.data = data
return True
async def notify_clients(self) -> None:
"""
Notify clients about events statuses in the queue periodically.
"""
while not self.stopped:
await asyncio.sleep(self.update_intervals)
if self.event_queue:
await self.broadcast_estimations()
async def broadcast_estimations(self) -> None:
estimation = self.get_estimation()
# Send all messages concurrently
await asyncio.gather(
*[
self.send_estimation(event, estimation, rank)
for rank, event in enumerate(self.event_queue)
]
)
async def send_estimation(
self, event: Event, estimation: Estimation, rank: int
) -> Estimation:
"""
Send estimation about ETA to the client.
Parameters:
event:
estimation:
rank:
"""
estimation.rank = rank
if self.avg_concurrent_process_time is not None:
estimation.rank_eta = (
estimation.rank * self.avg_concurrent_process_time
+ self.avg_process_time
)
if None not in self.active_jobs:
# Add estimated amount of time for a thread to get empty
estimation.rank_eta += self.avg_concurrent_process_time
client_awake = await self.send_message(event, estimation.dict())
if not client_awake:
await self.clean_event(event)
return estimation
def update_estimation(self, duration: float) -> None:
"""
Update estimation by last x element's average duration.
Parameters:
duration:
"""
self.duration_history_total += duration
self.duration_history_count += 1
self.avg_process_time = (
self.duration_history_total / self.duration_history_count
)
self.avg_concurrent_process_time = self.avg_process_time / min(
self.max_thread_count, self.duration_history_count
)
self.queue_duration = self.avg_concurrent_process_time * len(self.event_queue)
def get_estimation(self) -> Estimation:
return Estimation(
queue_size=len(self.event_queue),
avg_event_process_time=self.avg_process_time,
avg_event_concurrent_process_time=self.avg_concurrent_process_time,
queue_eta=self.queue_duration,
)
def get_request_params(self, websocket: fastapi.WebSocket) -> Dict[str, Any]:
return {
"url": str(websocket.url),
"headers": dict(websocket.headers),
"query_params": dict(websocket.query_params),
"path_params": dict(websocket.path_params),
"client": dict(host=websocket.client.host, port=websocket.client.port), # type: ignore
}
async def call_prediction(self, events: List[Event], batch: bool):
data = events[0].data
assert data is not None, "No event data"
token = events[0].token
data.event_id = events[0]._id if not batch else None
try:
data.request = self.get_request_params(events[0].websocket)
except ValueError:
pass
if batch:
data.data = list(zip(*[event.data.data for event in events if event.data]))
data.request = [
self.get_request_params(event.websocket)
for event in events
if event.data
]
data.batched = True
response = await AsyncRequest(
method=AsyncRequest.Method.POST,
url=f"{self.server_path}api/predict",
json=dict(data),
headers={"Authorization": f"Bearer {self.access_token}"},
cookies={"access-token": token} if token is not None else None,
client=self.queue_client,
)
return response
async def process_events(self, events: List[Event], batch: bool) -> None:
awake_events: List[Event] = []
try:
for event in events:
client_awake = await self.gather_event_data(event)
if client_awake:
client_awake = await self.send_message(
event, {"msg": "process_starts"}
)
if client_awake:
awake_events.append(event)
if not awake_events:
return
begin_time = time.time()
response = await self.call_prediction(awake_events, batch)
if response.has_exception:
for event in awake_events:
await self.send_message(
event,
{
"msg": "process_completed",
"output": {"error": str(response.exception)},
"success": False,
},
)
elif response.json.get("is_generating", False):
old_response = response
while response.json.get("is_generating", False):
# Python 3.7 doesn't have named tasks.
# In order to determine if a task was cancelled, we
# ping the websocket to see if it was closed mid-iteration.
if sys.version_info < (3, 8):
is_alive = await self.send_message(event, {"msg": "alive?"})
if not is_alive:
return
old_response = response
open_ws = []
for event in awake_events:
open = await self.send_message(
event,
{
"msg": "process_generating",
"output": old_response.json,
"success": old_response.status == 200,
},
)
open_ws.append(open)
awake_events = [
e for e, is_open in zip(awake_events, open_ws) if is_open
]
if not awake_events:
return
response = await self.call_prediction(awake_events, batch)
for event in awake_events:
if response.status != 200:
relevant_response = response
else:
relevant_response = old_response
await self.send_message(
event,
{
"msg": "process_completed",
"output": relevant_response.json,
"success": relevant_response.status == 200,
},
)
else:
output = copy.deepcopy(response.json)
for e, event in enumerate(awake_events):
if batch and "data" in output:
output["data"] = list(zip(*response.json.get("data")))[e]
await self.send_message(
event,
{
"msg": "process_completed",
"output": output,
"success": response.status == 200,
},
)
end_time = time.time()
if response.status == 200:
self.update_estimation(end_time - begin_time)
finally:
for event in awake_events:
try:
await event.disconnect()
except Exception:
pass
self.active_jobs[self.active_jobs.index(events)] = None
for event in awake_events:
await self.clean_event(event)
# Always reset the state of the iterator
# If the job finished successfully, this has no effect
# If the job is cancelled, this will enable future runs
# to start "from scratch"
await self.reset_iterators(event.session_hash, event.fn_index)
async def send_message(self, event, data: Dict, timeout: float | int = 1) -> bool:
try:
await asyncio.wait_for(
event.websocket.send_json(data=data), timeout=timeout
)
return True
except:
await self.clean_event(event)
return False
async def get_message(self, event, timeout=5) -> Tuple[PredictBody | None, bool]:
try:
data = await asyncio.wait_for(
event.websocket.receive_json(), timeout=timeout
)
return PredictBody(**data), True
except AsyncTimeOutError:
await self.clean_event(event)
return None, False
async def reset_iterators(self, session_hash: str, fn_index: int):
await AsyncRequest(
method=AsyncRequest.Method.POST,
url=f"{self.server_path}reset",
json={
"session_hash": session_hash,
"fn_index": fn_index,
},
client=self.queue_client,
)